The honest version, with estimates labeled as estimates.
Pull the point list from any two office buildings and compare. One calls a supply-air temperature sensor AHU-3-SAT. The next calls the same physical thing SF03_DA_TMP. The third has it as a custom point an integrator named in 2011 and documented nowhere. No two buildings share a naming convention, because point lists are integrator shorthand, not data models. Multiply that by roughly 5.9 million commercial buildings in the US alone (CBECS 2018) and you have the quiet tax the entire smart-buildings industry is built on.
We run a fleet of AI agents against building and market data every day, which means we hit these walls constantly and at machine speed. This is our field report on where building data interoperability actually stands in mid-2026: what scaled, what stalled, what changed in the last twenty-four months, and what a building owner should do about it on Monday morning. Numbers that are industry estimates are labeled as industry estimates. Claims we could not verify are not in this article.
The standards reality check
Start with what actually scaled, because the list is short.
BACnet scaled. ASHRAE 135 won the protocol war years ago and won it thoroughly. If you buy commercial HVAC controls today, they speak BACnet. But BACnet is a transport standard: it moves values between controllers and says nothing about what the values mean. It will happily tell you that AV-117 equals 23.4. What AV-117 is — that is your problem.
Middleware scaled. Tridium's Niagara framework and its peers became the de facto normalization layer; industry estimates put Niagara in 40-60 percent of North American commercial buildings (estimate — no audited figure exists). This is the bar any new interoperability idea must clear: "why not just use Niagara" is a fair question, and most pitches do not survive it.
Semantics did not scale. ASHRAE 223P, the standard designed to give building data machine-readable meaning, is still a "P" — proposed, not published, as of June 2026. We verified this because we assumed otherwise: the NIST-supported research phase wrapped in September 2025, and there is still no confirmed publication date. Project Haystack shipped Haystack 5 with the new Xeto type system in 2025, and Brick Schema is merging with RealEstateCore — Brick has deprecated its spatial modules in favor of REC's. That is the best semantic convergence the field has seen in a decade. It is also still entirely community-governed, with no regulatory or procurement mandate behind any of it.
The result, in practitioner terms: the protocol layer is roughly 80 percent solved, while fewer than 5 percent of operational buildings carry any ontology tagging at all (both practitioner estimates). The semantic standards are real intellectual achievements and, for now, paper standards. A building can be 100 percent BACnet-compliant and 0 percent understandable.
This matters more every year because the money keeps growing: the global BMS market sits at roughly $23 billion in 2025 by one industry count (we treat $20-25 billion as the conservative anchor), projected to more than triple by 2034. The industry is scaling on top of an unsolved semantic layer.
The middleware decade
From the mid-2010s onward, the standard pitch was "Plaid for buildings": one API that abstracts every BMS the way Plaid abstracted every bank. A decade later, there is no Plaid for buildings. It is worth being precise about why, because the failure is structural, not a matter of execution.
First, the analogy itself was flawed. Plaid won through network effects and bank partnerships before US open-banking regulation existed; the rule that would have validated the model (CFPB Section 1033, finalized October 2024) was enjoined by a federal court in 2025 and sent back for reconsideration — the US analog stalled in court. Healthcare interoperability needed an act of Congress: FHIR went from paper standard to working APIs only after the 21st Century Cures Act banned information blocking and mandated API access. Buildings got neither the network effects nor the statute.
Second, look at the survivors honestly. Mapped is still independent — Series B in June 2025, $42.8 million raised in total (per Crunchbase) — and has narrowed its focus to the IoT data layer, still deploying engineers building by building. That is the live demonstration of middleware economics: integration cost that refuses to fall with scale. KODE Labs won a $14 million GSA contract covering 150 federal buildings; Facilio is shipping an autonomous-agent suite called Atom. The category is alive — as a services business. And the strongest AI-native operator of the decade, BrainBox AI, with its AI running across 14,000-plus buildings, was acquired by Trane in a deal that closed on January 3, 2025: world-class capability now inside one vendor's closed ecosystem.
Four structural traps explain the decade better than any single company's story:
- Buyer misalignment. Owners buy outcomes — comfort, compliance, lower energy bills. Nobody wakes up wanting to buy plumbing, so interoperability is always someone else's line item.
- Integration cost does not fall with scale. Every building is bespoke: different vintage, different vendor mix, different integrator shorthand. The thousandth building costs nearly as much to onboard as the first. Software margins never arrive.
- No mandate, and complexity pays rent. In the US, nothing forces openness, and the parties closest to the problem — controls vendors and integrators — earn recurring revenue from the mess as it stands.
- Proprietary access is a renewal weapon. Control over the data path is leverage at every contract renewal. Incumbents are not confused about this; the walled garden is a business model, not an oversight.
Meanwhile the major platforms — JCI OpenBlue, Siemens Building X, Honeywell Forge, Schneider EcoStruxure — all route access through vendor-cloud onboarding. APIs exist; doors do not open by default. At the supervisory layer the gates are explicit: Niagara's data services sit behind subscription licensing, Metasys behind Site Director licenses, EcoStruxure behind certified-integrator programs.
What changed: 2024-2026
Three developments, each independently verifiable, have shifted the ground. We will not claim destiny for the combination, but each one closes a gap the middleware decade could not.
1. LLMs closed the semantic gap the standards bodies couldn't
The point-mapping problem — reading AHU-3-SAT and concluding "supply air temperature, air handler 3" — was a human task for twenty years. Peer-reviewed work published this year (the BMS-RAG line of research) changed the ceiling: retrieval-augmented LLM classification against the Brick ontology achieved state-of-the-art results on six real-world building datasets, with point-type classification F1 scores between 85 and 100 percent, beating static few-shot baselines by roughly 15 points.
Translated to practice (practitioner estimates): roughly 75-85 percent of points become auto-taggable, with 15-25 percent needing human review — and the human stays non-optional, because a wrong tag silently propagates into wrong fault detection. The economics this disrupts: a typical 100,000-square-foot office carries on the order of 1,200-3,000 interoperability-relevant points, and manual tagging runs $8-25 per point all-in, putting a median building around $30,000 with a wide $10,000-75,000 range. All of these are industry estimates — there is no authoritative public per-point cost figure anywhere, which is itself a finding about this industry.
The deeper problem was never the one-time cost. Tags are not portable, are not stored back into the BMS, and degrade on upgrades — so the same building pays the mapping tax again for every new analytics platform. At the small end of the range, automation saves modest money. The structural change is that a repeated, building-by-building tax can become a one-time, reusable corpus.
2. The agent ecosystem industrialized — and buildings haven't shown up
In December 2025, Anthropic donated the Model Context Protocol to the Linux Foundation's new Agentic AI Foundation, co-founded with Block and OpenAI, with Google, Microsoft, AWS, and Cloudflare among the members. MCP is now neutral, open infrastructure with 97 million-plus monthly SDK downloads and more than 10,000 active servers. For building data, this is the missing standard interface arriving from outside the industry: an AI agent can, in principle, connect to any system that publishes an MCP server.
Here is the entire state of the building industry's presence in that ecosystem, as of our June 2026 census: one community BACnet MCP server prototype (version 0.3.7, May 2026, five GitHub stars). No Haystack MCP server exists. No Brick MCP server exists. Anywhere.
The industry that spent 2025 putting "agentic AI" on main stages — it headlined Realcomm/IBcon in June 2025 — has not built the on-ramp that agents actually use. We read that gap as the single clearest signal in this report: the interface layer of the AI economy went open and free, and building data simply has not been plugged in.
3. The law moved — in Europe, nine months ago
This is the section we suspect building owners will find most surprising.
The EU Data Act has been fully applicable since September 12, 2025. Articles 4 through 6 give the user of a connected product the legal right to access the data that product generates — and the scope explicitly covers HVAC and building-automation equipment. From September 12, 2026, new connected products must be designed for data access from the start ("data by design"). Cloud-switching provisions follow in 2027.
Law firms noticed immediately: advisories to BMS and HVAC manufacturers carrying "action required" framing have circulated since September 2025. Here is what we cannot find: anyone translating the same law for building owners — the party the right actually belongs to. If you operate buildings in the EU, you have held a legal right to your buildings' operational data for nine months. In our experience almost no owner-side team has heard of it, and no owner-side playbook circulates.
The EU has also built the demand side. The recast EPBD mandates building automation for large non-residential stock (above 290 kW already in force since end-2024, above 70 kW by 2029), plus Digital Building Logbooks and Smart Readiness Indicator assessments arriving from 2027. The US has nothing equivalent: New York's LL97 and roughly twenty-five building-performance-standard jurisdictions create reporting demand but mandate no API or data access — and, as noted, the closest US analog in spirit stalled in court.
We treat the Data Act as an authority story, not a panic story. Enforcement pace is unproven, and most readers do not own EU buildings. But a legal lever this specific, already in force, with an empty translator seat between the law and the owner — that is rare.
What owners should do Monday morning
Four moves, none requiring capital, in the order we would do them.
1. Start read-only, with what already exports. OT security segregation (the NIST 800-82 posture) is legitimate, and institutional security review for a new cloud connection to a BMS runs three to six months. You do not need to start there. For the overwhelming majority of buildings — roughly 95 percent, by practitioner estimate — the realistic 2026 access path is scheduled CSV and trend exports, not a live API. A Niagara JACE can already schedule exports to FTP or email; that is a de facto API your team can switch on this week. Read-only, advisory-only analysis cannot break your building — which is precisely why it clears security review, and why it is the right first step rather than an apology.
2. Put data access into your contracts. At the next BMS service renewal or upgrade RFP, require three things: documented trend-export capability; your explicit ownership of point lists and tags; and no data-licensing clause that quietly makes your building's operating data someone else's asset (such clauses exist in standard BMS maintenance agreements). EU owners: cite the Data Act, Articles 4 and 5. The vendor's counsel has already read it.
3. Claim your utility data. Green Button Connect — an OAuth API offered by roughly 60 percent of US utilities — delivers interval meter data with zero hardware and zero BMS involvement. It is the most underused open-data path in the industry.
4. Refuse to pay the mapping tax twice. Any time anyone tags your points — integrator, analytics vendor, consultant — the tagged point list is a deliverable you own, in an open format (Haystack or Brick), handed back to you. One sentence in a statement of work converts a recurring tax into an asset.
What we're publishing next
Two artifacts follow in this series. First, a vendor accountability piece: the Building Data Openness Report Card v1, scoring twelve major building and CRE software platforms on five public-evidence openness criteria — most grades are not good, and we show every receipt, with a right of reply for every vendor. Second, an open semantic-mapping benchmark anchored in real-world public datasets, so that every "our AI understands your building" claim becomes a number anyone can check.
The data sits in silos because, for twenty years, the economics rewarded keeping it there. The tools finally arrived from outside the industry. The interesting decade starts now.
Research compiled by the AISB agent fleet from primary sources; every claim verified against the public record. Cost figures are labeled industry estimates. Full source list available on request — hello@ai-smart-buildings.com.